AuthorsHanyu Zhao, Yuqian Feng, Zhenyu Song et al.
TL;DR
LifeFuse-Mem uses lifecycle-aware subspace partition and protected state fusion to cut overwrite on Hard Attribution Anti-Overwrite from 42.28% to 36.29% and 39.73% to 30.61%.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Temporary overwrite occurs when temporary context is stored in the same compact state as permanent facts, silently altering future decisions.
LifeFuse-Mem targets long-running LLM agents whose persistent policies, like “confirm before external actions”, get overwritten by local instructions such as “skip confirmation this time”, degrading behavioral consistency and trust.
HOW IT WORKS
LifeFuse-Mem combines Lifecycle-Aware Subspace Partition, Permanent Routing and Subspace Writing, and Protected State Fusion and Lifecycle Readout to separate stable and transient knowledge.
Think of LifeFuse-Mem like splitting RAM into a protected region for long-term configuration and a scratchpad for temporary variables, then reading from each depending on query lifecycle.
This lifecycle-aware design lets LifeFuse-Mem answer permanent queries from preserved state while still following temporary overwrites for local behavior, something a plain context window or single shared memory state cannot do.
DIAGRAM
This diagram shows how LifeFuse-Mem processes Phase A permanent writes and Phase B temporary overwrites, then answers permanent and temporary queries.
DIAGRAM
This diagram shows the episodic write-then-read training loop and evaluation setup for LifeFuse-Mem on Hard Attribution Anti-Overwrite and public benchmarks.
PROCESS
Lifecycle-Aware Subspace Partition
LifeFuse-Mem learns an orthogonal basis Rl and Ql to split rank eight memory into plastic subspace A and stable subspace B, enabling lifecycle-specific control.
Permanent Routing and Subspace Writing
LifeFuse-Mem predicts a permanent-route weight ρt and scales stable rows so permanent tokens write strongly into B while temporary tokens mainly affect A.
Episodic Learning Objective
LifeFuse-Mem uses write-then-read training with LCE, Lroute, Lorth, Lhc, Lret, and Lsp to learn memory updates that preserve permanent facts under temporary overwrites.
Protected State Fusion and Lifecycle Readout
LifeFuse-Mem constructs Sstable by combining plastic rows from SAB and stable rows from SA, then fuses scores sA and sstable with weight α for permanent queries.
KEY CONTRIBUTIONS
Hard Attribution Anti-Overwrite benchmark
LifeFuse-Mem introduces Hard Attribution Anti-Overwrite with 1,000 episodes, six conflict keys, and four overwrite rounds per episode to directly measure temporary overwrite.
Lifecycle-Aware Subspace Partition
LifeFuse-Mem learns an orthogonal rank eight basis with rA = 4 plastic and rB = 4 stable dimensions, regularized by Lorth to keep lifecycles disentangled inside one memory state.
Protected State Fusion and Lifecycle Readout
LifeFuse-Mem fuses Phase A and protected stable-row states with weight α = 0.70, improving acquisition-controlled retention to 63.71% and 69.39% while lowering overwrite.
RESULTS
Ret.
63.71%
+6.0 over δ-Mem on Qwen3-4B
Ret.
69.39%
+9.12 over δ-Mem on SmolLM3-3B
Ovr.
36.29%
-6.0 overwrite vs δ-Mem on Qwen3-4B
Ovr.
30.61%
-9.12 overwrite vs δ-Mem on SmolLM3-3B
On the Hard Attribution Anti-Overwrite benchmark, which tests whether permanent facts survive conflicting temporary writes, LifeFuse-Mem improves acquisition-controlled retention and reduces overwrite relative to δ-Mem on both Qwen3-4B and SmolLM3-3B, showing that lifecycle-aware memory can selectively protect durable knowledge.
BENCHMARK
On the Hard Attribution Anti-Overwrite benchmark, which tests whether permanent facts survive conflicting temporary writes, LifeFuse-Mem improves acquisition-controlled retention and reduces overwrite relative to δ-Mem on both Qwen3-4B and SmolLM3-3B, showing that lifecycle-aware memory can selectively protect durable knowledge.
BENCHMARK
Retention (Ret.) on conflict-permanent facts after Phase B, conditioned on correct acquisition after Phase A.
BENCHMARK
Overall LoCoMo percentage score across multi-hop, temporal, open-domain, and single-hop categories.
KEY INSIGHT
LifeFuse-Mem cuts overwrite from 42.28% to 36.29% and 39.73% to 30.61% without materially changing temporary query accuracy.
This is surprising because one might expect stronger protection of permanent facts to blunt adaptation to Phase B, but LifeFuse-Mem preserves temporary behavior while selectively shielding stable knowledge.
WHY IT MATTERS
LifeFuse-Mem shows that compact online memory can be lifecycle-aware, keeping long-term policies intact while still honoring temporary instructions.
Builders can now design persistent agents whose internal state resists silent drift from transient contexts, enabling safer long-horizon behavior and more trustworthy personalization.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…